Fargo Ratings, Different Starting #s?

[sbpoolleague]

I understand that Fargo Ratings are not used for 14.1 (straight pool). If this is true, why not? Is there a universally accepted rating system for 14.1?

Thanks,
Will

Mike Page correct me if I'm wrong....
Fargo Ratings are based on whole games won or lost, not points won in a particular game. Games such as 14.1 and Banks do not have clearly defined winning objectives. You can play to any predefined score. Even if a score was mandated for Fargo entry for these games (eg. race to 25 points), the games usually take a lot longer to play than a single game of 8, 9, or 10 ball. For that reason, the game of One Pocket also can't be used. The Fargo Rate formula needs the varying game units to be roughly equivalent.

Oh, and the Fargo Rate formula, just like APA's Equiliizer formula, Weight Watchers points formula, etc. are all proprietary formulas. That's how they make their money.

 

[Will Maynard]

Fargo ratings are basically used for 8ball. 9ball. and 10ball.
One Pocket and 14.1 are completely different games. I have seen some handicap formulas used in One Pocket and we used to use a local 14.1 handicap system but I'm not aware of any National 14.1 ratings.

Thanks for the feedback. I would love to see a 14.1 rating system at the pro level to show relative playing strength using results from high level tournaments like the World 14.1 Championship and now the US Open. Not tons of games available, but if tournament results were applied from say the last 10 years I think it would be enough data to have a reasonable rating. 14.1 matches are generally 100, 150, or 200 point games, but I don't think that would skew the rating results. To me, a match is a match, regardless of the target score.

Your thoughts?

Thanks,
Will

 

[Will Maynard]

Mike Page correct me if I'm wrong....
Fargo Ratings are based on whole games won or lost, not points won in a particular game. Games such as 14.1 and Banks do not have clearly defined winning objectives. You can play to any predefined score. Even if a score was mandated for Fargo entry for these games (eg. race to 25 points), the games usually take a lot longer to play than a single game of 8, 9, or 10 ball. For that reason, the game of One Pocket also can't be used. The Fargo Rate formula needs the varying game units to be roughly equivalent.

Oh, and the Fargo Rate formula, just like APA's Equiliizer formula, Weight Watchers points formula, etc. are all proprietary formulas. That's how they make their money.

Thanks for your thoughts.

I found a March, 2010 Fargo Ratings paper by Mike Page showing Fargo Ratings equations in detail. If the current broadly-applied Fargo Rating approach is proprietary then it must have changed since the 2010 paper. MIke Page.....please comment.

I am interested in how the equations and approach differs from the ELO rating system used in chess. ELO does a pretty good job at rating individuals in head-to-head matches which I have applied to 14.1 players and it seems to be quite accurate at predicting outcomes. I would rather use Fargo if it is better.

Thanks for your interest,
Will

 

[AtLarge]

... I found a March, 2010 Fargo Ratings paper by Mike Page showing Fargo Ratings equations in detail. If the current broadly-applied Fargo Rating approach is proprietary then it must have changed since the 2010 paper. MIke Page.....please comment.

I am interested in how the equations and approach differs from the ELO rating system used in chess. ELO does a pretty good job at rating individuals in head-to-head matches which I have applied to 14.1 players and it seems to be quite accurate at predicting outcomes. I would rather use Fargo if it is better. ...

While we wait for Mike's response, here's something he wrote 4½ months ago on why he hadn't posted a formula:

"There is not a formula to give. It is a mathematical optimization process. You can think of all the ratings of every player in the world as variables. And you can think of specific game/match results among those players as outcomes that are either likely or unlikely depending upon on the values of the ratings. Then you can imagine there is a set of ratings for which every match in the system coming out the way it did is most likely. This is a technique in statistical inference called maximum liklihood. The results are well defined. We just can't point to a simple formula."

And here are a couple comments from Mike about FargoRate vs. ELO and chess:

"Yes, we talked back in 2010. What we were doing back then--and what we talked about--was an ELO scheme that was basically a modest improvement on what FIDE does in chess. What we do now --what we call the global optimization--is very different. So hang in there. I think you will like what you see...." 12/29/15

"Regardless, all those--FIDE, Glick, Glicko-2... are sequential approaches to updating the ratings. We used to do that. Now we have taken a big leap to the global optimization..." 12/2/15

So it looks like the equations you found from 2010 are not the current approach.

Edit -- Oh, and welcome aboard, Will.

 

[BmoreMoney]

All Paul's games were attached to one record. In general multiple entries result from merging databases, and we want to be orderly about merging records because we don't want to break important connections. But these will be merged into one.

Wow Mike; I admire you and feel soooo bad for you at the same time lol! I know I feel soooo soooo bad for the TD of a lowly local handicapped tourney with 32people *****in and crying about ratings, dont even want to think about what its like with 65,000 lol. You're doing a great job though bro, keep on keeping on!

 

[pocket]

I play in a BCAPL and have a few hundred games played over a few seasons. Also played in a league tourney, yet I only have a 44 game robustness. Is it something the league has to submit or should it be automatically posted?

It has stayed that way for as long as I have checked.

Just curious.

 

[Will Maynard]

While we wait for Mike's response, here's something he wrote 4½ months ago on why he hadn't posted a formula:

"There is not a formula to give. It is a mathematical optimization process. You can think of all the ratings of every player in the world as variables. And you can think of specific game/match results among those players as outcomes that are either likely or unlikely depending upon on the values of the ratings. Then you can imagine there is a set of ratings for which every match in the system coming out the way it did is most likely. This is a technique in statistical inference called maximum liklihood. The results are well defined. We just can't point to a simple formula."

And here are a couple comments from Mike about FargoRate vs. ELO and chess:

"Yes, we talked back in 2010. What we were doing back then--and what we talked about--was an ELO scheme that was basically a modest improvement on what FIDE does in chess. What we do now --what we call the global optimization--is very different. So hang in there. I think you will like what you see...." 12/29/15

"Regardless, all those--FIDE, Glick, Glicko-2... are sequential approaches to updating the ratings. We used to do that. Now we have taken a big leap to the global optimization..." 12/2/15

So it looks like the equations you found from 2010 are not the current approach.

Edit -- Oh, and welcome aboard, Will.

Thanks very much for your comprehensive and enlightening response. I have much to learn about the process and I will continue to pursue.

I feel strongly that 14.1 deserves a global rating system. Although 14.1 has not recently enjoyed the status it once did in the Greenleaf/Mosconi era it is a wonderful and complex game that when played at the highest level is amazing to watch. With the resurrection of the 14.1 US Open and the continued 14.1 World Tournament et al, hopefully we see a much deserved return to popularity Straight Pool.

 

[mikepage]

What are the formulas for computing a Fargo Rating? How is a Fargo Rating calculated for someone without a rating who plays one game with a player with a rating?

Thanks

Fargo Ratings follow a maximum likelihood approach. So if you play 3 games against a known opponent (and that is all you have played) and win 2 of them, FargoRate will assume you will win 2 of the next 3, and 20 of the next 30, and so forth.

If you play a single game against a rated opponent and lose, the system can't distinguish you from a rock, and you will have a minus infinity rating.

If you win the game, the system can't distinguish you from a god, and you will have a plus infinity rating.

 

[mikepage]

Thanks for the feedback. I would love to see a 14.1 rating system at the pro level to show relative playing strength using results from high level tournaments like the World 14.1 Championship and now the US Open. Not tons of games available, but if tournament results were applied from say the last 10 years I think it would be enough data to have a reasonable rating. 14.1 matches are generally 100, 150, or 200 point games, but I don't think that would skew the rating results. To me, a match is a match, regardless of the target score.

Your thoughts?

Thanks,
Will

The Fargo Rating approach works just fine for straight pool using a point as the basic scoring unit.

The difficulty that SBpoolleague refers to is how to combine straight-pool results with, say, 9-ball results in computing a composite rating. This is an issue we call the run-length issue. If A beats B 40 to 20 in 9-ball and then later loses to B 20 to 40 in straight pool, most of us would conclude A is likely a better player, even though their overall "score" is 60 to 60. There is more information in a game of 9-ball than there is in a single point of straight pool. You can apply a multiplier between the two scoring units that is computed either from run-length statistics or--with a lot of data--determined empirically.

A couple years ago I ran a straight-pool league in Fargo. All the players had established ratings based on 8-ball only. Of course everybody said it is apples and oranges, etc. But when we look at the independent end-of-season 14.1 ratings and compare them to the 8-Ball ratings the overall correspondence is pretty good.

 

[BRussell]

Fargo Ratings follow a maximum likelihood approach. So if you play 3 games against a known opponent (and that is all you have played) and win 2 of them, FargoRate will assume you will win 2 of the next 3, and 20 of the next 30, and so forth.

That seems wrong to me. Every prediction approach (e.g. regression or Bayesian) predicts something closer to 50-50 from such a small sample like 3 games. So from winning 2 of 3 games, rather than predicting 20 of 30, any reasonable prediction would predict something more like, say, 18/30 wins, and for 300 games something more like 160/300. Anthing else assumes that even a very small sample is a perfect predictor.

 

[mikepage]

That seems wrong to me. Every prediction approach (e.g. regression or Bayesian) predicts something closer to 50-50 from such a small sample like 3 games. So from winning 2 of 3 games, rather than predicting 20 of 30, any reasonable prediction would predict something more like, say, 18/30 wins, and for 300 games something more like 160/300. Anthing else assumes that even a very small sample is a perfect predictor.

responding to Mr. Russell -- this will sound like gibberish to most others.

Assuming without data a player is not a rock and is not a god--aka incorporation of a bayesean prior probablility--sounds like a very reasonable thing to do. And we have done a MAP (Maximum a posteriori) approach with all our data using a beta-distribution prior (conjugate to bernoulli). This does what you suggest above and can be shown to be completely equivalent to assuming an unrated player has already played and split some games against a hypothetical average player (could be 20 games at 10 to 10 or or 2 games at 1 to 1 or one game at .5 to .5) or whatever.


This approach works just fine with a large and thoroughly coupled group. It basically has no effect on anybody who has played many games, and players with very few games have ratings that may not mean much but at least make them look human.

But there is a serious flaw with this approach for the global optimization--in particular for multiple large weakly coupled groups--that has caused us to reject it. Because of this we are frequentists through and through.

 

[Bob Jewett]

That seems wrong to me. Every prediction approach (e.g. regression or Bayesian) predicts something closer to 50-50 from such a small sample like 3 games. So from winning 2 of 3 games, rather than predicting 20 of 30, any reasonable prediction would predict something more like, say, 18/30 wins, and for 300 games something more like 160/300. Anthing else assumes that even a very small sample is a perfect predictor.

Hmmm... I think it's clear though that the match-win ratio between the players that gives the highest probability of a 2-1 result is a 2:1 win ratio (p=2/3, q=1/3).

 

[mikepage]

Hmmm... I think it's clear though that the match-win ratio between the players that gives the highest probability of a 2-1 result is a 2:1 win ratio (p=2/3, q=1/3).

Yes, and that is what we do.

What Mr Russell is saying is the we don't really go into it completely clueless.

Suppose SVB goes to a tournament and plays three games against an unknown player from an obscure region of an obscure country, and Shane loses 2 of the games.

At this point the MLE approach--what we do--says our best guess is that this guy is 100 points above SVB and is the best player in the world by far. We also say we have very low confidence in this conclusion.

But any reasonable observer would say the chance is way higher that this is not the best player in the world and is likely not as good as Shane.

Many approaches take this kind of thing into account--an automatic bias toward the mean. We don't, and we don't for very good reason

 

[Papa Red]

How will this system work with guys playing in multiable leagues. Lets say a guy plays in a handicapped league and has an 80% win avg. and in the other league with A+, AA, & AAA with a 45% win avg.

 

[mikepage]

How will this system work with guys playing in multiable leagues. Lets say a guy plays in a handicapped league and has an 80% win avg. and in the other league with A+, AA, & AAA with a 45% win avg.

A player can establish a Fargo Rating--the same Fargo Rating--by playing in just the easy league or just the hard league--or both. All that is required is SOME players from those leagues being connected to the outside world.

 

[pocket]

I play in a BCAPL and have a few hundred games played over a few seasons. Also played in a league tourney, yet I only have a 44 game robustness. Is it something the league has to submit or should it be automatically posted?

It has stayed that way for as long as I have checked.

Just curious.

Mike if you get a chance could you address my question? Thank you sir.

 

[one stroke]

Does the BCA enter thier league Data into Fargo


1

 

[aquaurchin99]

I just created an Fargo account. How do I get my own statistics on there? I see some of the strong players local to me have ratings. I do not have anything when i search myself. How is stuff added?

 

[BRussell]

Mike if you get a chance could you address my question? Thank you sir.

I'm in a similar situation - I've been doing a BCA league for a couple of years, and quite a few local tournaments, but I have no games in the system. My friends who have gone to nationals do have games in the system.

 

[Will Maynard]

The Fargo Rating approach works just fine for straight pool using a point as the basic scoring unit.

The difficulty that SBpoolleague refers to is how to combine straight-pool results with, say, 9-ball results in computing a composite rating. This is an issue we call the run-length issue. If A beats B 40 to 20 in 9-ball and then later loses to B 20 to 40 in straight pool, most of us would conclude A is likely a better player, even though their overall "score" is 60 to 60. There is more information in a game of 9-ball than there is in a single point of straight pool. You can apply a multiplier between the two scoring units that is computed either from run-length statistics or--with a lot of data--determined empirically.

A couple years ago I ran a straight-pool league in Fargo. All the players had established ratings based on 8-ball only. Of course everybody said it is apples and oranges, etc. But when we look at the independent end-of-season 14.1 ratings and compare them to the 8-Ball ratings the overall correspondence is pretty good.

I'm not surprised at the similar results between straight pool ratings and 8-ball since these 2 games require somewhat similar skills. Straight pool ratings vs a rotation-based game ratings may show more differences as different skills are required for these 2 games.

Why not have separate ratings for straight pool and rotation-based games?

Thanks for your comments, Mike. This is a very interesting topic.

Will